# Yoke Create > Skill factory for the yoke plugin. Full pipeline for creating a new skill: task analysis, design with a mermaid diagram, implementation of SKILL.md and agents, quality validation, documentation creation, updating README and CLAUDE.md. Used when the user writes "create a skill", "new skill", "add a skill", "make a skill", "skill for". - Skill: `yokeloop/yoke-create` (Agent Skill) - Install (CLI): `npx skillmds@latest add yokeloop/yoke-create` - Raw SKILL.md: https://api.skillmd.com/api/skills/yokeloop/yoke-create/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: yokeloop (https://skillmd.com/u/yokeloop) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/yokeloop/yoke-create --- # Skill factory You are the orchestrator for creating new skills for the yoke plugin. Coordinate the work through TodoWrite and sub-agents. Use AskUserQuestion for clarifications and confirmations. Work without pauses between phases, except for Phase 3 (plan confirmation). --- ## Input `$ARGUMENTS` — skill description: what it should do, usage examples, ticket URL, or path to a file with the description. If `$ARGUMENTS` is empty — ask via AskUserQuestion: "Describe what the new skill should do." --- ## Pipeline 8 phases. Create a TodoWrite at the start: ``` [ ] Preflight: verify the project root [ ] Analyze: understand the task and context [ ] Design: design the skill with a mermaid diagram [ ] Confirm: agree on the plan with the user [ ] Implement: create SKILL.md, agents, reference [ ] Validate: check prose quality and structure [ ] Integrate: docs, README, CLAUDE.md [ ] Complete: final summary ``` --- ## Phase 0 — Preflight Verify that we are in the yoke project root: ```bash test -f .claude-plugin/plugin.json && test -d skills/ ``` If not → report: "Run from the yoke project root." Exit. --- ## Phase 1 — Analyze Run 2 agents in parallel (model: sonnet, subagent_type: general-purpose). ### Agent 1 — Task analysis Prompt: > Analyze the user's request for creating a new skill for the yoke plugin. > > User request: `<$ARGUMENTS>` > > Determine: > 1. **GOAL** — what the skill should do (1-2 sentences) > 2. **TRIGGERS** — which phrases activate it (at least 5 trigger phrases in Russian and English) > 3. **INPUT** — what it accepts through $ARGUMENTS > 4. **OUTPUT** — what it produces (artifacts, files, actions) > 5. **PHASES** — preliminary decomposition into phases (3-7 phases) > 6. **AGENTS_NEEDED** — whether sub-agents are needed, and if so — which roles > 7. **REFERENCES_NEEDED** — whether reference files with templates/formats are needed > 8. **COMPLEXITY** — simple (1-2 phases, no agents) | medium (3-5 phases, 1-3 agents) | complex (5+ phases, 3+ agents) > > If the request contains a URL or a file path — read it and extract requirements. > > Return structured output with the fields above. ### Agent 2 — Analysis of existing skills Prompt: > Analyze all existing skills in the `skills/` directory of the yoke project to ensure consistency of the new skill. > > For each `skills/*/SKILL.md`: > 1. Read the frontmatter (name, description) > 2. Determine the pattern: number of phases, number of agents, models used (haiku/sonnet/opus) > 3. Note stylistic conventions: language, header format, phrase templates > > Also read: > - One docs file (`docs/review.md`) as a documentation template > - README.md — the "Skills" section as a template for the new entry > - CLAUDE.md — the "Implemented skills" section > > Return: > ``` > SKILL_COUNT: > COMMON_PHASES: > AGENT_MODELS: > STYLE_CONVENTIONS: > DOC_TEMPLATE: > README_TEMPLATE: > CLAUDE_MD_TEMPLATE: > ``` Wait for both agents. Merge the results: TASK_ANALYSIS + CONVENTIONS. TodoWrite: mark "Analyze" as done. --- ## Phase 2 — Design Based on the analysis, design the skill. Act yourself (without agents). ### 2a. Choose the name Name = kebab-case, English, 1-2 words. Propose 2-3 variants via AskUserQuestion. ### 2b. Design the architecture Determine: - **Phases** — list of phases with names and descriptions - **Agents** — for each agent: name, role, model (haiku for data gathering, sonnet for analysis, opus for implementation), tools - **Reference files** — templates and formats if needed - **Input/Output** — what it accepts and what it produces ### 2c. Mermaid flow diagram Create a mermaid diagram of the skill's flow: ```mermaid flowchart TD A[Input: $ARGUMENTS] --> B[Phase 1: ...] B --> C{Decision} C -->|Option 1| D[Phase 2: ...] C -->|Option 2| E[Output] D --> F[Phase 3: ...] ``` Include: phases, decision points (AskUserQuestion), agents (as sub-processes), artifacts. ### 2d. Compose the plan Combine into a plan: ``` Skill: / — Phases: 1. — [agent: , model: ] 2. — ... Agents: - (model) — ... Reference: - .md — ... Artifacts: Mermaid: ``` TodoWrite: mark "Design" as done. --- ## Phase 3 — Confirm Show the plan to the user (including the mermaid diagram). AskUserQuestion: - **Approve and implement** (Recommended) - **Refine the plan** — the user describes what to change → return to Phase 2 - **Cancel** → exit Maximum 3 refinement cycles. After the third — only "Approve" or "Cancel". TodoWrite: mark "Confirm" as done. --- ## Phase 4 — Implement Create all skill files. Act sequentially. ### 4a. Directory structure ```bash mkdir -p skills//agents skills//reference ``` Create `agents/` and `reference/` only if they are needed by the plan. ### 4b. SKILL.md Create `skills//SKILL.md` via the Write tool. Follow yoke conventions: - **Frontmatter**: `name` and `description` (description in third person, at least 5 trigger phrases) - **Heading**: `# ` - **Role**: "You are the orchestrator." + brief description - **Input**: a `## Input` section describing `$ARGUMENTS` - **Phases**: `## Phase N — ` with transitions - **Rules**: a `## Rules` section at the end - **Language**: match the ticket/input language, or follow the project-level definition in CLAUDE.md / AGENTS.md. - **Imperative**: "Run", "Create", "Verify" (not "You must") - **TodoWrite**: mark each phase ### 4c. Agents For each agent create `skills//agents/.md`: ```yaml --- name: description: >- tools: model: color: --- ``` Model selection conventions: - **haiku**: data gathering, formatting, writing files (fast operations) - **sonnet**: analysis, research, validation (balanced) - **opus**: code implementation, complex logic (quality is critical) Color conventions: - **cyan**: git operations and infrastructure - **blue**: core implementation - **green**: polishing and formatting - **yellow**: warnings and alerts ### 4d. Reference files Create `skills//reference/.md` for templates, formats, and guides. ### 4e. Examples (optional) If the skill produces artifacts — create `skills//examples/` with examples of output files. TodoWrite: mark "Implement" as done. --- ## Phase 5 — Validate Run 2 agents in parallel (model: sonnet). ### Agent 3 — Prose check (elements-of-style) Prompt: > Check the text quality in SKILL.md and the agent files of the new skill `skills//`. > > Read all .md files in `skills//`. > > Check against the rules: > 1. **Active voice** — "The agent gathers data" not "Data is gathered by the agent" > 2. **Positive statements** — "Use X" not "Don't forget to use X" > 3. **Concrete language** — no "various", "appropriate", "certain" > 4. **Extra words** — no "in general", "essentially", "basically" > 5. **Brevity** — can it be shortened without losing meaning > 6. **Imperative** — "Run" not "You need to run" > > Return: > ``` > FILES_CHECKED: > ISSUES_COUNT: > ISSUES: > - FILE: | LINE: | RULE: | FIX: > ``` ### Agent 4 — Structure validation (skill-development) Prompt: > Validate the structure of the new skill `skills//` against best practices. > > Read all files in `skills//`. > > Checks: > 1. **Frontmatter**: SKILL.md has name and description > 2. **Name**: matches the directory name > 3. **Description**: third person, at least 3 trigger phrases, not vague > 4. **Agent frontmatter**: each agents/*.md has name, description, tools, model, color > 5. **Agent references**: all agents mentioned in SKILL.md exist in agents/ > 6. **Reference integrity**: all reference/ files mentioned in SKILL.md exist > 7. **Body size**: SKILL.md < 500 lines (recommendation) > 8. **Phases**: phases are numbered, have names > 9. **TodoWrite**: mentioned for progress tracking > 10. **Rules section**: a "Rules" section at the end of SKILL.md > > Return: > ``` > CHECKS_PASSED: > CHECKS_FAILED: > ISSUES: > - CHECK: | DETAIL: > ``` Wait for both agents. If issues > 0 → show the list, fix the problems via the Edit tool, report what was fixed. TodoWrite: mark "Validate" as done. --- ## Phase 6 — Integrate ### 6a. Documentation Create `docs/.md` from the template: ```markdown # Skill / <1-2 sentences: what the skill does> ## Input ``` /yoke: ``` ## Phases | Phase | Name | What happens | | ----- | ------- | ------------ | | 1 | **...** | ... | ## Output ## Sub-agents | Agent | Model | Role | | ----- | ----- | ---- | | `...` | ... | ... | ## Example ## Relations ``` ### 6b. Sync docs Run `/sync-docs` to regenerate the per-skill MDX page under `site/src/content/docs/skills/`, the catalog table between `` markers in `README.md`, and the bullet list between the same markers in `CLAUDE.md`. On failure → name the file sync-docs reports and halt. ### 6c. Format ```bash pnpm run format ``` TodoWrite: mark "Integrate" as done. --- ## Phase 7 — Complete Print the summary: ``` Skill / created Files: skills//SKILL.md skills//agents/.md skills//agents/.md skills//reference/.md docs/.md Updated: README.md — added a / section CLAUDE.md — added to Implemented skills ``` AskUserQuestion: - **Commit via /yoke:gca** (Recommended) - **Test the skill** — the user tests it and returns with feedback - **Finish** → exit If "Test" is chosen: wait for feedback, fix the issues via the Edit tool, show the updated summary. TodoWrite: mark "Complete" as done. --- ## Rules - Language: match the ticket/input language, or follow the project-level definition in CLAUDE.md / AGENTS.md. Commit messages and frontmatter `name` — in English. - Files and directories — kebab-case. - Description in SKILL.md and agents — third person with trigger phrases. - Phase 1 and Phase 5 agents — read-only, do not modify files. - Maximum 3 plan refinement cycles in Phase 3. - Do not create empty directories (agents/, reference/) if they are not needed. - On error — show the problem and propose a solution. - Mark each phase via TodoWrite immediately upon completion.